sched/cache: Keep nr_pref_llc_running in the runnable domain, to fix LLC mis-scheduling buggrafted

alb_break_llc() decides whether to break LLC preference during active
load balance. It does so by testing that every runnable fair task on the
source rq prefers its LLC:

	env->src_rq->nr_pref_llc_running == env->src_rq->cfs.h_nr_runnable

But the two counters cover different sets. nr_pref_llc_running is updated
in account_llc_enqueue()/account_llc_dequeue(), next to cfs_rq->nr_queued,
so it follows queued tasks. h_nr_runnable is updated in set_delayed()/
clear_delayed() and drops delay-dequeued tasks.

So under DELAY_DEQUEUE, a preferring task that goes to sleep stays counted
in nr_pref_llc_running while h_nr_runnable falls. The equality then breaks,
alb_break_llc() returns false, and active balance is free to pull a task
off its preferred LLC. Active balance only moves runnable tasks, and this
is the only LLC check it consults: once the stopper runs, LBF_ACTIVE_LB
skips the per-task test in can_migrate_task(). The runnable set is the one
we want.

Fix it on the counter side. A task should be counted in
nr_pref_llc_running exactly while it is both queued on its preferred LLC
(pref_llc_queued) and runnable (!sched_delayed). Define that membership
once in task_pref_llc_runnable(), and adjust the counter only through
pref_llc_running_inc()/pref_llc_running_dec() from the four sites that
change either input: account_llc_enqueue(), account_llc_dequeue(),
set_delayed() and clear_delayed(). Gating every update on the same
predicate keeps the delay, wake and dequeue paths from double-counting
or underflowing; see the comments at those sites for the ordering.

nr_llc_running and sd->llc_counts are not touched and stay on queued
semantics.

Fixes: 714059f79ff0 ("sched/cache: Handle moving single tasks to/from their preferred LLC")
Closes: https://lore.kernel.org/lkml/20260827135000.735138-1-zhanxusheng@xiaomi.com/
Reported-by: Zhan Xusheng <zhanxusheng@xiaomi.com>
Suggested-by: Chen Yu <yu.c.chen@intel.com>
Signed-off-by: Tim Chen <tim.c.chen@linux.intel.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Reviewed-by: Kayra Cizmeci <kayracizmeci@gmail.com>
Cc: <stable@kernel.org> # v7.2.x
Link: https://patch.msgid.link/06af61afedac32e6477f57feb4d658f6c411c3af.1790035273.git.tim.c.chen@linux.intel.com

1 files changed, 132 insertions, 0 deletions

diff --git a/Documentation/driver-api/rfkill.rst b/Documentation/driver-api/rfkill.rst
new file mode 100644
index 000000000..7d3684e81
--- /dev/null
+++ b/Documentation/driver-api/rfkill.rst
@@ -0,0 +1,132 @@
+===============================
+rfkill - RF kill switch support
+===============================
+
+
+.. contents::
+   :depth: 2
+
+Introduction
+============
+
+The rfkill subsystem provides a generic interface for disabling any radio
+transmitter in the system. When a transmitter is blocked, it shall not
+radiate any power.
+
+The subsystem also provides the ability to react on button presses and
+disable all transmitters of a certain type (or all). This is intended for
+situations where transmitters need to be turned off, for example on
+aircraft.
+
+The rfkill subsystem has a concept of "hard" and "soft" block, which
+differ little in their meaning (block == transmitters off) but rather in
+whether they can be changed or not:
+
+ - hard block
+	read-only radio block that cannot be overridden by software
+
+ - soft block
+	writable radio block (need not be readable) that is set by
+        the system software.
+
+The rfkill subsystem has two parameters, rfkill.default_state and
+rfkill.master_switch_mode, which are documented in
+admin-guide/kernel-parameters.rst.
+
+
+Implementation details
+======================
+
+The rfkill subsystem is composed of three main components:
+
+ * the rfkill core,
+ * the deprecated rfkill-input module (an input layer handler, being
+   replaced by userspace policy code) and
+ * the rfkill drivers.
+
+The rfkill core provides API for kernel drivers to register their radio
+transmitter with the kernel, methods for turning it on and off, and letting
+the system know about hardware-disabled states that may be implemented on
+the device.
+
+The rfkill core code also notifies userspace of state changes, and provides
+ways for userspace to query the current states. See the "Userspace support"
+section below.
+
+When the device is hard-blocked (either by a call to rfkill_set_hw_state()
+or from query_hw_block), set_block() will be invoked for additional software
+block, but drivers can ignore the method call since they can use the return
+value of the function rfkill_set_hw_state() to sync the software state
+instead of keeping track of calls to set_block(). In fact, drivers should
+use the return value of rfkill_set_hw_state() unless the hardware actually
+keeps track of soft and hard block separately.
+
+
+Kernel API
+==========
+
+Drivers for radio transmitters normally implement an rfkill driver.
+
+Platform drivers might implement input devices if the rfkill button is just
+that, a button. If that button influences the hardware then you need to
+implement an rfkill driver instead. This also applies if the platform provides
+a way to turn on/off the transmitter(s).
+
+For some platforms, it is possible that the hardware state changes during
+suspend/hibernation, in which case it will be necessary to update the rfkill
+core with the current state at resume time.
+
+To create an rfkill driver, driver's Kconfig needs to have::
+
+	depends on RFKILL || !RFKILL
+
+to ensure the driver cannot be built-in when rfkill is modular. The !RFKILL
+case allows the driver to be built when rfkill is not configured, in which
+case all rfkill API can still be used but will be provided by static inlines
+which compile to almost nothing.
+
+Calling rfkill_set_hw_state() when a state change happens is required from
+rfkill drivers that control devices that can be hard-blocked unless they also
+assign the poll_hw_block() callback (then the rfkill core will poll the
+device). Don't do this unless you cannot get the event in any other way.
+
+rfkill provides per-switch LED triggers, which can be used to drive LEDs
+according to the switch state (LED_FULL when blocked, LED_OFF otherwise).
+
+
+Userspace support
+=================
+
+The recommended userspace interface to use is /dev/rfkill, which is a misc
+character device that allows userspace to obtain and set the state of rfkill
+devices and sets of devices. It also notifies userspace about device addition
+and removal. The API is a simple read/write API that is defined in
+linux/rfkill.h, with one ioctl that allows turning off the deprecated input
+handler in the kernel for the transition period.
+
+Except for the one ioctl, communication with the kernel is done via read()
+and write() of instances of 'struct rfkill_event'. In this structure, the
+soft and hard block are properly separated (unlike sysfs, see below) and
+userspace is able to get a consistent snapshot of all rfkill devices in the
+system. Also, it is possible to switch all rfkill drivers (or all drivers of
+a specified type) into a state which also updates the default state for
+hotplugged devices.
+
+After an application opens /dev/rfkill, it can read the current state of all
+devices. Changes can be obtained by either polling the descriptor for
+hotplug or state change events or by listening for uevents emitted by the
+rfkill core framework.
+
+Additionally, each rfkill device is registered in sysfs and emits uevents.
+
+rfkill devices issue uevents (with an action of "change"), with the following
+environment variables set::
+
+	RFKILL_NAME
+	RFKILL_STATE
+	RFKILL_TYPE
+
+The content of these variables corresponds to the "name", "state" and
+"type" sysfs files explained above.
+
+For further details consult Documentation/ABI/stable/sysfs-class-rfkill.